romsi-hate-speech 1.0.1

Detect Romanized Sinhala hate speech using mBERT.

Romanized Sinhala Hate Speech Detection

Detect hate speech in Romanized Sinhala text using a fine-tuned deep learning model based on mBERT.
This project includes:
✅ Training code
✅ Inference code
✅ Pip-installable package
✅ REST API (FastAPI)
✅ CLI tool

---

📚 About

This project was developed as part of my MSc research to address hate speech detection in Romanized Sinhala, commonly used in Sri Lankan social media.

The model is trained on the SOLD dataset, fine-tuned on mBERT, and exposes predictions through a REST API, a CLI, and as a reusable Python package.

---

🚀 Features

• Fine-tuned multilingual transformer for Romanized Sinhala text.
• Inference available via:
  • Python package (romsi_hate_speech)
  • CLI command: romsi-detect
  • REST API (FastAPI server)
• Training pipeline to reproduce experiments.
• MIT licensed and open source.

---

🗂️ Project Structure

romanized_hate_speech_detection/
├── romsi_hate_speech/         # Inference package (pip-installable)
│   ├── __init__.py
│   ├── predictor.py
│   ├── api.py
│   ├── cli.py
├── training/                  # Training and evaluation code
│   ├── trainer.py
│   ├── evaluator.py
│   ├── data_loader.py
│   └── ...
├── models/                    # Saved models
├── data/                      # Datasets and preprocessing scripts
├── README.md                  # This file
├── setup.py                   # Packaging metadata
├── requirements.txt
├── .gitignore
├── LICENSE

---

🔷 Installation

You can install the inference package locally:

pip install .

Or (once published):

pip install romsi-hate-speech

---

🧪 Usage

🐍 Python

from romsi_hate_speech import Predictor

predictor = Predictor(model_path="sakunchamikara/romsi-hate-speech")
label, confidence = predictor.predict("meka thamai mage msc research project eka")
print(label, confidence)

💻 CLI

romsi-detect "meka thamai mage msc research project eka"

or for multiple Texts

from romsi_hate_speech.predictor import Predictor

predictor = Predictor(model_path="sakunchamikara/romsi-hate-speech")
texts = [
    "patta horekta yahapalanayen adhyaksha thanathurak",
    "marila palayan balla"
]
results = predictor.predict(texts)

for r in results:
    print(f'"{r["text"]}" → {r["label"]} (confidence: {r["confidence"]})')

🌐 REST API

Run the API server:

uvicorn romsi_hate_speech.api:app --reload

Then open: http://127.0.0.1:8000/docs

Or POST to /predict:

{
  "texts": ["patta horekta yahapalanayen"]
}

---

📈 Training

To reproduce training:

python training/model_trainer.py

You can configure hyperparameters in training/config.py.

---

⚖️ License

This project is licensed under the MIT License.

---

👤 Author

• Sakun Chamikara
• MSc Research Project, 2025

---

🌐 Links

• HuggingFace Model
• PyPI Package
• Chrome Extension
• Fly.io API

---